Number of conversational contacts

Number of conversational contacts outside the household

Conversational contacts by age

Conversational contacts outside household by age

By age/sex

By age/sex - non-hh contacts

Comparison - all conversational contacts

# A tibble: 16 x 6
# Groups:   ego_age [4]
   ego_age  alter_age  bics    fb ratio frac_decrease
   <chr>    <chr>     <dbl> <dbl> <dbl>         <dbl>
 1 [25,35)  [25,35)   1.09  7.16  0.152         0.848
 2 [25,35)  [35,45)   0.5   2.36  0.212         0.788
 3 [25,35)  [45,65)   0.451 1.25  0.361         0.639
 4 [25,35)  [65,100]  0.112 0.173 0.646         0.354
 5 [35,45)  [25,35)   0.513 3.29  0.156         0.844
 6 [35,45)  [35,45)   1.29  5.87  0.220         0.780
 7 [35,45)  [45,65)   0.367 1.70  0.215         0.785
 8 [35,45)  [65,100]  0.202 0.528 0.383         0.617
 9 [45,65)  [25,35)   0.317 2.23  0.142         0.858
10 [45,65)  [35,45)   0.411 3.10  0.132         0.868
11 [45,65)  [45,65)   0.943 3.77  0.250         0.750
12 [45,65)  [65,100]  0.319 0.755 0.422         0.578
13 [65,100] [25,35)   0.222 0.737 0.301         0.699
14 [65,100] [35,45)   0.322 2.14  0.150         0.850
15 [65,100] [45,65)   0.451 2.00  0.225         0.775
16 [65,100] [65,100]  0.791 2.17  0.364         0.636

Relationships

Relationships, non-household contacts

Locations

Locations - non-household contacts

Contact durations - by relationship

Contact durations - by respondent age

Awareness

Concern

Behavior change

Number of interviews

Conversational contacts

Conversational contacts outside of household

Overview

To help summarize patterns in the contact survey data, we fit a negative binomial model, accounting for the right-censoring of reported contacts at 10. These models show relationships among people who have completed the survey; have not been adjusted in any way for sampling. We fit these models using the brms package in R.

We modeled the expected log number of reported contacts as a function of age group, city, gender, and household size. The plots below show posterior means and 95% credible intervals for the estimated coefficients. Estimated coefficients greater than 0 imply that the predictor is associated with higher reported numbers of contacts, while estimated coefficients less than 0 imply that the predictor is associated with lower reported numbers of contacts.

There are two models: one for total number of contacts, and one for the number of non-household contacts.

Negative-binomial model for household contacts

Negative-binomial model for non-household contacts

Technical details

We fit negative binomial models to summarize the patterns in the data, and to account for right-censoring of reported contacts at 10.

That is, we model the observed reported number of conversational contacts, \(y_i\), as

\[ \begin{aligned} y_i &\sim \text{Poisson}(\lambda_i)\\ \lambda_i &\sim \text{Gamma}() \end{aligned} \]

We fit models of the form

$$ \begin{aligned}

i &= + {male[i]} + {age[i]} + {city[i]} + _{hhsize} _i, \end{aligned} $$

where \(\alpha\) is an intercept; \(i\) indexes survey respondents; \(\beta_{male[i]}\) is the coefficient on a dummy variable for whether or not respondent is male; \(\beta_{city[i]}\) is a dummy variable for the sample \(i\) comes from (national, or one of the city samples; reference group is the national sample); \(\text{hh}_i\) is \(i\)’s reported household size; and \(\beta_{hhsize}\) is the estimated coefficient on household size.

Age/sex

Race/ethnicity

Household size

Microdata

We plan to produce a version of the data with no identifying information publicly available as soon as we can. If you are a disease modeler who urgently needs to see the microdata, please reach out to us by email.

The estimated mixing matrices are reproduced as tables below. Note that these are the crude estimates, and have not had a symmetry constraint enforced.

All contact mixing estimates

Non-household mixing estimates